A feature-driven acquisition strategy using scale-invariant descriptors for deep active learning in preclinical CT segmentation

Warren, Faye, Paisey, Stephen ORCID: https://orcid.org/0000-0002-2274-3708, Spezi, Emiliano ORCID: https://orcid.org/0000-0002-1452-8813, Lai, Yu-kun ORCID: https://orcid.org/0000-0002-2094-5680 and Smith, Rhodri 2025. A feature-driven acquisition strategy using scale-invariant descriptors for deep active learning in preclinical CT segmentation. Presented at: 29th Medical Image Understanding and Analysis Conference, Leeds, UK, 15-17 July 2025. Published in: Ali, S., Hogg, D. C. and Peckham, M. eds. Proceedings of MIUA. Annual Conference on Medical Image Understanding and Analysis , vol.15918 Springer, pp. 129-145. 10.1007/978-3-031-98694-9_10

Full text not available from this repository.

Abstract

Accurate medical image segmentation is essential for clinical decision-making, and while deep learning (DL) models have substantially advanced automated segmentation, their performance depends heavily on large, well-annotated datasets, which are often costly and time-consuming to produce, especially in specialised or emerging imaging scenarios. Active learning (AL) addresses this limitation by iteratively selecting the most informative samples for annotation, ensuring the training dataset is diverse and representative of the overall data distribution to maximise model performance with fewer annotations. This work proposes a novel acquisition function that, for the first time, integrates Scale-Invariant Feature Transform (SIFT) descriptors into AL. By combining entropy-based uncertainty sampling with SIFT-based keypoint matching, we introduce a hybrid strategy—SIFT-Entropy—that leverages both learned uncertainty and feature-driven diversity to enhance training efficiency. We validate SIFT-Entropy by comparing its performance against uncertainty and random sampling in a preclinical whole-body mouse Computer Tomography (CT) segmentation task using a Dense VNet. The experiments assess segmentation performance across multiple labels using the Dice Similarity Coefficient (DSC). SIFT-Entropy accelerates model performance improvements, achieving full dataset performance with 61% fewer annotated samples. Furthermore, interquartile range (IQR) analysis confirms the consistency of these performance gains. These findings highlight the potential of hybrid acquisition functions in AL for medical image segmentation, offering a pathway toward data-efficient DL models in resource-constrained environments.

Item Type:	Conference or Workshop Item (Paper)
Date Type:	Publication
Status:	Published
Schools:	Schools > Engineering Schools > Computer Science & Informatics Schools > Medicine
Publisher:	Springer
ISBN:	978-3-031-98693-2
Last Modified:	18 Aug 2025 13:59
URI:	https://orca.cardiff.ac.uk/id/eprint/180081

Actions (repository staff only)

Edit Item